Fuzzy System for Adapted Generation of ...

3 downloads 0 Views 349KB Size Report
[12] M. Person and J. Bergensten, “Minecraft,” 2009, Mojang. [13] J. C. Rose, D. G. Souza, A. L. Rossito, T. M. S. Rose, 'Aquisição de leitura após história de ...
Fuzzy System for Adapted Generation of Educational Tasks for Children with Reading Disabilities Adalberto Bosco C. Pereira1 and Leonardo B. Marques2 and Dionne C. Monteiro1 and Gilberto Nerino de Souza1 and Clay Palmeira da Silva1. Abstract.1 This work proposes a computational approach that searches to use resources of Artificial Intelligence in Education (AIED) to aid teachers, psychologists and educationalists in the learning process of reading. This approach aims at generating teaching tasks which can be accordingly adapted to the individual needs of each student. While the tasks are being executed, a Machine Learning (ML) system will collect and process data to allow an analysis of the student learning process for each individual word in reading and writing. The fuzzy system will propose an appropriate task based on the data collected by the ML. The output of fuzzy system is an adapted task that will motivate the children in execution of the tasks.

1

INTRODUCTION

Digital games get an important space in the lives of children, teenagers and adults, and are now one of the fastest growing sectors in the media and entertainment industry. In addition, games have been combined with a wide range of fields such as unusual simulations, Games for Health, Game-Based Learning (GBL) [1] Artificial Intelligence in Education (AIED) [2] and many others. Good learning requires teachers and students to combine their efforts. In the case of students, their interest in studying is of fundamental importance [3] to ensure they can realize their potential to the maximum. With regard to teachers, it can be concluded that any attempt to improve students’ achievements must be based on the acquisition of an effective teaching behavior. [4]. In other words, they should give students appropriate guidance. For several years, some games have incorporated automatic generation to create levels, missions and space that are designed to increase the lifetime of their games. There are games that have a form of challenge called ‘endless’, where the player is exposed to challenges at levels that are generated indefinitely until the game is lost. This approach is adopted for several areas in games and in many features such as 2D textures, 3D models, music, levels, story, mission and so on. These studies are of increasing importance in the process of developing computer games [5]. The purpose of this work is to test a solution of Fuzzy Logic applied to the current task of teaching reading and writing in a digital game that is adaptive to the individual needs of each player. That is, if the new tasks of education are suited to the pre-existing Laboratory of Applied Artificial Intelligence – Institute of Exact and Natural Sciences – Federal University of Para, Brazil, email: [email protected], [email protected], [email protected], [email protected] 2 Laboratory for the Study of Human Behavior – Federal University of Sao Carlos, Brazil, email: [email protected] 1

literacy skills of the student, this means that this task must not be either too easy or too difficult. The data collected during the execution of the teaching tasks will be pre-processed and analyzed with the aid of a Machine Learning (ML), which will provide data to Fuzzy System to evaluate and consider the best choice for proposing a task. The output of the Fuzzy System is determined by the data required for the generation of the task, which in turn will be transformed into an adapted level of the game.

2

RELATED WORK

The “Gerenciador de Ensino Individualizado por Computador” (GEIC - An Individualized Education Manager for Computers) [6] approaches the problem of ensuring the dynamic generation of content for educational use, by providing software for programming procedures based on teaching tasks of choice. It allows the creation of Teaching Units that combine various teaching tasks, and represent discrete attempts to provide choice tasks. The objective of this work is to aid teachers in teaching reading and writing to children with learning difficulties. Further details of this program will be given in Section Three of this paper. M.A. Azevedo [7] discusses the methods of teaching, which although of a satisfactory standard, still have different efficiencies for each student. This is discussed in examining the computational and educational resources and their degree of adaptability to the individual needs of each student. Studies in the field of AIED [2] by B.du Benedict, state that the individualization of teaching instruction can be effective and is regarded by the author as the “Holy Grail of AIED”. This paper compares the educational differences in AI systems (AIED) with conventional educational systems used in the classroom or traditional methods of Computer-Assisted Instruction (CAI). The learning program called GEIC [6], referred to above, was transformed into a game called ALE-RPG [8]. However, it should be stressed that the structured progress that the game kept in the GEIC, proved to be a little static. The progression of the player through the teaching tasks does not allow a fully customized advance to be made, since it is conditioned by the need to learn all the words of each teaching unit. . The non-acquisition of any word component of the teaching units delays the teaching progress of the next units. There are games like Diablo [9], Torchlight [10], Spore [11] and MineCraft [12] that use automatic levels of generation when the player starts a new game, but keeps the missions and game objectives distributed randomly on the map that was generated.

However, all this randomness follows ‘brute force’ algorithms, without any predetermined scale of difficulty. The approach of D. Dormans [5] investigates strategies to generate levels of action-adventure games that are divided into two individual structures, so that they generate missions first and then spaces. The different types of generative grammar are analyzed in a search for the one that best fits. L. Xiangfeng [1] uses Fuzzy Cognitive Maps to design GameBased Learning (GBL). The goal is to use the Hebbian learning rule to increase learning capacity by employing the game data and Unbalance Degree to establish the lack of prior knowledge.

3

TEACHING PROGRAM

Since the 1980s, researchers in Brazil in the area of behavior psychology have been refining a program to help children with a previous history of school failure, learn to read. This procedure, called “Aprendendo a Ler e a Escrever em Pequenos Passos” (ALEPP - Learning to Read and Write in Small Steps) [13], is mainly concerned with detecting and overcoming problems that are found in children who have literacy difficulties, in an efficient manner. It also provides tasks that teach the basic components of reading in a personalized way.. The GEIC is a remote software that allows the ALEPP curriculum [13] to be applied. But it is not yet adaptive enough, because it is composed of static tasks, grouped into Teaching Units that have been determined by specialists previously. The Matching to Sample (MTS) procedure is used for teaching reading relations programmed into ALEPP [13] through GEIC [6], Figure 1. This procedure taught the relations between printed words, pictures and dictated words. In this task, one stimulus (the sample) must be matched to the correct comparison. Each task contains n comparisons, ranging from one to three, where one acts as a stimulus by referring to the sample on top of screen. Finally, the last feature is the definition of the words that correspond to the model stimulus and stimulus choices. The tasks are subdivided into types of combinations of stimulus such as AB, BC and CB. In Figure 1 (b) the sample stimulus is the sound that corresponds to the correct alternative.

Tasks allow you to create relationships between the stimuli of different modalities. They are relations that are taught between the dictated words (A), representative figure of this word (B) and the printed word (C). The tasks that establish these relations are listed as tasks of type AB (dictated word-picture), CB (printed wordpicture) and BC (picture-printed word).

4

PERFORMANCE ASSESSMENT IN READING

In the current ALEPP assessment system, the GEIC automatically load the teaching tasks assessments [15]. The GEIC can be set individually, at the end of each teaching session, if the student has repeated the last block of tasks or advanced to the next block. This software allows a minimum number of correct answers to be set in a session, and uses these successes as a criterion for advancement. However it is not a fine evaluation in terms of the learning words. The final session assessment is unable to identify the different kinds of errors for each word taught. The ideal situation is evaluating the correctness of each word. For example, in attempts of the word "pipe" as sample it appears that the student learned the word, how he knows this word, and then a task is set with a focus on learning how to read or write, so that the student can learn to read or write that word.

5

PROPOSAL

In adopting an approach to assist instructors, the general purpose of the project is initially aimed at creating one AIED to teach reading and writing and then incorporate it into a digital game. AIED is composed of two intelligent systems, ML and Fuzzy System, which will act together, as shown in the diagram in Figure 2. They represent the Machine Learning system that is not covered in this work, but developed separately in parallel. In the first stage of the project, are discussed strategies and methods for teachings reading are employed.

(b)

(a)

Figure 2. Macro view of project.

The ML is responsible for evaluating the student, deciding the level of expertise and supplying input data so the (c) Figure 1. MTS Task Model. AB (a) Type CB (b) and type BC (c).

Figure 3. Flow diagram of the proposal.

fuzzy system can generate the task that has been adapted. The fuzzy logic will be discussed in detail later. The system is a digital game that begins with an initial pretest of static tasks to generate the minimum data required to enable the Machine Learning to correctly analyze the student. After the information obtained during the gameplay, for each word that has been processed, a new task is generated by Fuzzy System and will be stored in the database; the level of the game is created by using the features of the task generated. The sequence of tasks follows that of Logical Teaching because for the specialists, there is a preferred order for each word being taught, and new words will enter the teaching tasks gradually, depending on the degree of literacy of the word. For each Teaching Unit there are fifteen words that have to be taught. The ML will decide the level of knowledge and if the student has learned every single word. As in the case of the execution of new ML generated tasks, they will always be calculated on the basis of this new information regarding the degree of knowledge of a particular word, as the player learns. The words that are determined as literacy appear less frequently, and when every word has been learned, the game is over. Figure 3 shows a diagram of the operation of the proposal.

end of the system, the output is the completion of the task, as shown in Figure 4.

6

All the data processed and generated by ML, are abstracted and normalized in three fuzzy sets corresponding to a numerical range, from 0 to 100%, and a degree of pertinence ranging from 0 to 100% [17] as shown in Figure 2. The fuzzy values represented in this approach, correspond to the descriptors for trapezoidal and triangular functions illustrated in Figure 5. In the present work all the fuzzy sets have these same values of classification.

FUZZY SYSTEM

The main objective of this paper is to make use of data generated by ML correctly in order to generate an adjusted task. Fuzzy Logic was chosen to carry this out for the following reasons: it has a distinct capacity to express the vagueness and uncertainty of the knowledge it represents [16], it is able to model a system close to logical grammatical rules, it ensures a better approximation to the knowledge of a specialist, through semantic representations and linguistic terms, and it operates by choosing few rules and working with imprecise terms [17]. The data provided by ML are fuzzified and separated into fuzzy groups. For each task feature, one fuzzy inference occurs. At the

Figure 4. Macro view of Fuzzy System.

6.1

Fuzzification

6.3

Fuzzy Inference

The fuzzy inference system employed in this work follows the Mamdani model [19], which corresponds to the algorithm of fuzzified information processing in accordance with linguistic rules [20] which are defined by the specialists and research studies referred to. Table 2 correlates the input variables with the output in logical terms and the "if-then" form in causal terms. In the initial testing, the following rules were obtained and are arranged in Table 3 (a), (b) and (c) as follows: "If a variable column 1 = X and variable column 2 = Y then variable column 3 = Z". Table 2. Variables of outputs. Figure 5. Graphical representation of fuzzy set partitioning for a linguistic variable, showing the vertical and horizontal axes referring to the degree of pertinence values and fuzzy respectively.

Output Variable DTT: Need for Task Type. DNC: Need for number of comparisons.

This Fuzzification is done individually for each input variable of a type Task, Number of Comparisons and Incorrect Words of a particular model of word, described in Table 1: Table 1. Input variables of the fuzzy system.

Input Variable PTT: Probability of hit with determined Task Type. TTT: Hit Rate of Task Type. PNC: Probability of hit with determined number of Comparisons. TNC: Hit Rate of number of Comparisons. PPI: Probability of hit with determined incorrect word. TPI: Hit Rate with determined Incorrect Word.

6.2

DPI: Need for incorrect word. Table 3. Rules: (a) type of task, (b) number of comparisons and (c) misspelled words.

PTT Low Low Low Moderate Moderate Moderate High High High

TTT Low Moderate High Low Moderate High Low Moderate High (a)

DTT Low Low Moderate Moderate High Moderate Moderate High Low

PNC Low Low Low Moderate Moderate Moderate High High High

TNC Low Moderate High Low Moderate High Low Moderate High (b)

DNC Low Low Moderate Moderate High Moderate Moderate High Low

PPI Low Low Low Moderate Moderate Moderate High High High

TPI Low Moderate High Low Moderate High Low Moderate High (c)

DPI Low Moderate Moderate Low High Moderate Moderate Low Low

Rule Sets

The architecture of the Fuzzy Logic, provided here, is mapped out in a set of rules reflecting the ideas of specialists about the Study of Human Behavior [18] project, as well as the interviews that were conducted for this study and the works outlined above. The set of rules for this work is designed to create the appropriate learning tasks in an efficient manner, and also aims at encouraging the students to ensure the game is kept enjoyable and stimulating. If the tasks that are generated are too difficult for the player and he starts to miss too much, the player will not learn and may be discouraged and lose interest in the game. On the other hand, if there is too little difficulty for a player, he may also become discouraged, and will not be able to play the game to its full potential, and thus be shortchanging the teaching process, and this will delay his learning. From this perspective, the aim is to strike a balance between degrees of difficulty for each student. The rules that are set (together with specialists) attempt to treat the input variables so that the choices of output variables, which are the characteristics of the new task, have an appropriate degree of difficulty. In addition, a further aspect of this logic is to ensure that the same feature does not appear too often and thus , prevents the tasks from becoming repetitive. Each task performed by the ML player will pass on updated data about the student’s progress, which means that the rules of the fuzzy system will always generate an appropriate task based on the updated data.

As an example that points to a rule set for the task type BC: “If the Probability of the Task Type BC is Low (pertinence 85%) and the Hit Rate of Task Type is High (63% pertinence) then the Need for Task Type BC is Medium (pertinence 74%)”, as shown in Figure 7.

the analysis by the ML. At the same time we must include AIED in the game that is under development. The game that is currently under development is being implemented in a Game Engine called Unity3D, because this tool allows one to generate various games for platforms such as PC, Xbox360, Web HTML, Flash, iOS and Android. The aim is to ensure that the end of the game runs on the Tablet with Android.

9

Figure 7. Application of fuzzy logic to the task type.

6.4

Decision-making

Finally, it should be decided which task type is better to choose. The best linguistic value of fuzzy set assigned as Highest is chosen. The above logic is applied to establish the number of incorrect comparisons, while only changing the choice of words. Instead of only one value, the algorithm that chooses the words returns, a list with n words, where n is previously defined by the fuzzy logic responsible for deciding the number of comparisons. There is no need for this work to defuzzify the output variables, because the choice is given by the highest degree of pertinence as stated previously, thus generating task features for the subsequent level of the game.

This work is part of a project that proposes the use of Intelligent Agents attached to computer games and targeted at the learning of reading and writing. This part of the system being researched aims at exploring the question of decision-making when there is uncertainty about the real need for a particular person to perform a given task type. The first results were obtained by processing data from the database project where there is a record of how the GEIC program was implemented. With this information it was possible to simulate results to allow the ML to convey on information to the Fuzzy System, and then it was validated in a satisfactory way by specialists in the field of psychology. However, it is still necessary to make adjustments to the variables, address other tasks, examine other fuzzy variables and also cover the tests directly in the computer games with the students running the system in real time. It is concluded that the project is viable with positive results encouraging us to continue expanding the research to able the game to teaching writing and improvement it.

REFERENCES [1] [2] [3] [4]

7

RESULTS [5]

The work is still in its initial phase; and to validate the operation and efficiency of the proposal, the Fuzzy System algorithms were implemented by means of the C # language. The ALEPP and GEIC teaching programs are currently being tested at some schools in the city of São Carlos, SP - Brazil. This database which has records of the teaching program for each child, was used to yield the ML results, and thus generated data for the initial tests in the fuzzy system. , These tests were conducted with the assistance and under the supervision of specialists, as well as on the basis of an analysis of the results. The system was validated by specialists, and usually led to the creation of a task that could to be the same task designed for a specialist.

8

FUTURE STUDIES

Despite obtaining satisfactory results, there is still a need to carry out tests in a complete game running in real time. Future developments will focus on employing the fuzzy system to generate writing tasks, and thus teach the player to read and write simultaneously. It is also expected that there can be an extension of

CONCLUSION

[6]

[7] [8]

[9] [10] [11] [12] [13]

[14]

L. Xiangfeng, W. Xiao and Z. Jun ‘Guided Game-Based Learning’, Published by IEEE Transactions on Learning Technologies, (2010). B.du Benedict. ‘What does the “AI” in AIED buy?’ Printed and published by IEE, Savoy Place, London WC2R 0BL, U.K., 1997. S. Tobias, “Interest, Prior Knowledge, and Learning,” Rev. of Educational Research,vol. 64, no. 1, pp. 37-54, 1994 J. Brophy, “Teacher Influences on Student Achievement,” Am. Psychologist,vol. 41, no. 10, pp. 1069-1077, 1986. D. Dormans and B. Sander, ‘Generating Missions and Spaces for Adaptable Play Experiences”, Published by IEEE Transactions on Computational Intelligence AI in Games, 2011. L. B. Marques, R. G. Meio, R. M. Maria, ‘Manual do Usuário de Programas de Ensino via GEIC’ - Volume 1: ‘Aprendendo a Ler e Escrever em Pequenos Passos’. São Carlos, 2011. M. A. Azevedo, M. L Marques, ‘Alfabetização hoje’. São Paulo: Cortez, 2001. E. S. Sarmanho, E. B. Sales, D. M. Cavalcante, L. B. Marques, ‘Um Jogo com Reconhecedor de Voz para o Ensino de Crianças com Dificuldade de Aprendizagem em Leitura e Escrita’. Published by Semish, 2011. E. Schaefer, D. Brevik, M. Schaefer, E. Sexton, and K. William, “Diablo,” 1996”, Blizzard North. T. Baldree, “Torchlight,” 2009, Runic Games. W.Wright, “Spore,” 2008, Maxis. M. Person and J. Bergensten, “Minecraft,” 2009, Mojang. J. C. Rose, D. G. Souza, A. L. Rossito, T. M. S. Rose, ‘Aquisição de leitura após história de fracasso escolar: equivalência de estímulos e generalização’. In: Psicologia: Teoria e Pesquisa, p.451-69. 1989. de Souza, D. G., de Rose, J. C., Faleiros, T. C., Bortoloti, R., Hanna, E. S., & McIlvane, W. J. Teaching Generative Reading Via Recombination of Minimal Textual Units: A Legacy of Verbal Behavior to Children in Brazil. Revista Internacional De Psicologia

[15]

[16]

[17]

[18] [19]

[20]

Y Terapia Psicologica - International Journal of Psychology and Psychological Therapy, 9(1), pp 19–44, 2009. T. S. Reis, D. G. Souza, J. C. Rose, ‘Avaliação de um programa para o ensino de leitura e escrita’. In: Estudos em Avaliação Educacional, 20, p.425-50. 2009. C. D. Pedro, J. O. Adriano, ‘Aprendizado de Regras Nebulosas em Tempo Real para Jogos Eletrônicos’. XI Brazilian Symposium of Multimedia Systems and Web. Games – II Brazilian Workshop of Games and Digital Entertainment, 2003. P. B. Moratori, M. V. Pedro, L. M. B. Manhaes, C. Lima, A. J. O. Cruz, E. B. Ferreira, and L. C. V. de Andrade. ‘Analysis of the Stability of a Fuzzy Control System Developed to Control a Simulated Robot,’ Fuzzy Systems, 2005. The 14th IEEE International Conference on Fuzzy, pp.726-730, 25-25 May 2005. L.B. MARQUES, ‘Variáveis Motivacionais no Ensino de Leitura: O jogo como recurso complementar’, ed. São Carlos: UFSCar, 2009. Mamdani, E.H., "Advances in the linguistic synthesis of fuzzy controllers," International Journal of Man-Machine Studies, Vol. 8, pp. 669-678, 1976. R. J. Timothy Fuzzy Logic with Engineering Applications, Third Edition", ISBN: 047074376X, Wi ey, 5, 117-148, 2010.